Retrieval Control

(Execution-Time Retrieval)

When an AI step runs:

• It retrieves current information from connected systems.
• Retrieval occurs at execution time via live API call.
• It does not rely on cached snapshots or model memory.

Retrieval may include:

• Unstructured documents (via ingestion pipelines).
• Structured operational records (via direct query).

Retrieval is:

• Scoped to configured connector surfaces.
• Bound to the workflow contract.
• Correlated to execution telemetry.

There is no global knowledge layer.
There is no cross-tenant memory.

---

Context Control

(Workflow-Scoped Context Persistence)

Each workflow maintains bounded execution state.

Context includes:

• Completed steps
• Prior decisions
• State transitions
• Intermediate data

Context is:

• Tenant-scoped
• Workflow-scoped
• Correlated by execution identifiers

An AI step operates on:

1. Current system state (Retrieval Control)
2. Accumulated workflow state (Context Control)

It does not access unrelated organizational memory.

---

Replay & Failure Model

Workflow execution is checkpointed at defined node boundaries.

If a failure occurs:

• The boundary is recorded.
• Replay resumes from the failed node.
• Completed steps are not re-executed.
• Workflow context does not reset.

External side effects remain subject to provider idempotency constraints.

There is:

• No distributed atomicity across third-party systems.
• No global exactly-once guarantee.

Failure is bounded and traceable.

---

Boundaries

• Retrieval applies only to connected systems.
• Structured retrieval currently uses direct query paths.
• Adapter-bound canonical drift containment is not assumed for structured retrieval at this stage.
• There is no global AI memory.
• This is not a conversational AI layer.

---